def test_generate_resnet_model():
architecture = generate_resnet_architecture(n_upsampling=3,
n_filter=64,
stride=2)
input_params = {'shape': [128, 128, 3]}
model = load_architecture(architecture, input_params)
model.summary()
expected_model = load_architecture(
load_json_config('tests/json/resnet_architecture.json'), input_params)
assert check_same_models(model, expected_model)
def test_generate_voxresnet_model():
architecture = generate_voxresnet_architecture(n_upsampling=3,
n_filter=64,
stride=2)
input_params = {'shape': [128, 128, 128, 3]}
model = load_architecture(architecture, input_params)
model.summary()
def test_generate_resnet_model_dropout():
architecture = generate_resnet_architecture(n_upsampling=4,
n_filter=48,
stride=2,
dropout_rate=0.1)
input_params = {'shape': [128, 128, 3]}
model = load_architecture(architecture, input_params)
model.summary()
def test_generate_densenet_model_3d():
architecture = generate_densenet_3d_architecture(n_upsampling=3,
n_filter=[4, 8, 12, 16],
dense_block=[2, 3, 4, 5],
stride=2)
input_params = {'shape': [128, 128, 128, 3]}
model = load_architecture(architecture, input_params)
model.summary()
def test_generate_densenet_model_batchnorm():
architecture = generate_densenet_2d_architecture(n_upsampling=4,
n_filter=48,
stride=2,
batchnorm=True)
input_params = {'shape': [128, 128, 3]}
model = load_architecture(architecture, input_params)
model.summary()
def test_generate_vnet_model_resize():
architecture = generate_vnet_architecture(n_upsampling=5,
n_filter=4,
stride=2)
input_params = {'shape': [129, 129, 129, 3]}
model = load_architecture(architecture, input_params)
assert np.all(model.input_shape == (None, 129, 129, 129, 3))
assert np.all(model.output_shape == (None, 129, 129, 129, 1))
def test_load_vnet_model_resize():
architecture = load_json_config(
read_file('tests/json/vnet_architecture.json')
)
input_params = {'shape': [129, 129, 129, 3]}
model = load_architecture(architecture, input_params)
assert np.all(model.input_shape == (None, 129, 129, 129, 3))
assert np.all(model.output_shape == (None, 129, 129, 129, 1))
def test_load_sequential_model():
architecture = load_json_config(
read_file('tests/json/sequential_architecture.json'))
input_params = {'shape': [32, 32]}
model = load_architecture(architecture, input_params)
input_layer = Input(shape=(32, 32))
flatten = Flatten()(input_layer)
dense = Dense(units=128, activation='relu')(flatten)
dropout = Dropout(rate=0.2)(dense)
output_layer = Dense(units=10, activation='softmax')(dropout)
expected_model = Model(inputs=input_layer, outputs=output_layer)
assert check_same_models(model, expected_model)
def test_generate_unet_model():
architecture = generate_unet_architecture(n_upsampling=5,
n_filter=4,
stride=2)
input_params = {'shape': [128, 128, 3]}
model = load_architecture(architecture, input_params)
model.summary()
def conv_layers(filters, pre_layer):
conv = BatchNormalization()(Conv2D(
filters,
kernel_size=3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(pre_layer))
return BatchNormalization()(Conv2D(
filters,
kernel_size=3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv))
input_layer = Input(shape=[128, 128, 3])
conv_1 = conv_layers(4, input_layer)
max_pool_1 = MaxPooling2D()(conv_1)
conv_2 = conv_layers(8, max_pool_1)
max_pool_2 = MaxPooling2D()(conv_2)
conv_3 = conv_layers(16, max_pool_2)
max_pool_3 = MaxPooling2D()(conv_3)
conv_4 = conv_layers(32, max_pool_3)
max_pool_4 = MaxPooling2D()(conv_4)
conv_5 = conv_layers(64, max_pool_4)
max_pool_5 = MaxPooling2D()(conv_5)
conv_t_kwargs = {
"kernel_size": 3,
"strides": 2,
"padding": "same",
"kernel_initializer": "he_normal"
}
conv_6 = conv_layers(128, max_pool_5)
conv_trans_1 = Conv2DTranspose(64, **conv_t_kwargs)(conv_6)
upconv_1 = conv_layers(64, concatenate([conv_5, conv_trans_1]))
conv_trans_2 = Conv2DTranspose(32, **conv_t_kwargs)(upconv_1)
upconv_2 = conv_layers(32, concatenate([conv_4, conv_trans_2]))
conv_trans_3 = Conv2DTranspose(16, **conv_t_kwargs)(upconv_2)
upconv_3 = conv_layers(16, concatenate([conv_3, conv_trans_3]))
conv_trans_4 = Conv2DTranspose(8, **conv_t_kwargs)(upconv_3)
upconv_4 = conv_layers(8, concatenate([conv_2, conv_trans_4]))
conv_trans_5 = Conv2DTranspose(4, **conv_t_kwargs)(upconv_4)
upconv_5 = conv_layers(4, concatenate([conv_1, conv_trans_5]))
output = Conv2D(1,
kernel_size=3,
activation='sigmoid',
padding="same",
kernel_initializer="he_normal")(upconv_5)
expected_model = Model(inputs=input_layer, outputs=output)
expected_model.summary()
assert check_same_models(model, expected_model)
def test_load_vnet_model():
architecture = load_json_config(
read_file('tests/json/vnet_architecture.json')
)
input_params = {'shape': [128, 128, 128, 3]}
model = load_architecture(architecture, input_params)
def conv_layers(filters, pre_layer):
conv = BatchNormalization()(
Conv3D(filters,
kernel_size=3,
activation='relu',
padding='same')(pre_layer))
return BatchNormalization()(
Conv3D(filters,
kernel_size=3,
activation='relu',
padding='same')(conv))
input_layer = Input(shape=[128, 128, 128, 3])
conv_1 = conv_layers(4, input_layer)
max_pool_1 = MaxPooling3D()(conv_1)
conv_2 = conv_layers(8, max_pool_1)
max_pool_2 = MaxPooling3D()(conv_2)
conv_3 = conv_layers(16, max_pool_2)
max_pool_3 = MaxPooling3D()(conv_3)
conv_4 = conv_layers(32, max_pool_3)
max_pool_4 = MaxPooling3D()(conv_4)
conv_5 = conv_layers(64, max_pool_4)
max_pool_5 = MaxPooling3D()(conv_5)
conv_t_kwargs = {"kernel_size": 3,
"strides": [
2,
2,
2
],
"padding": "same"}
conv_6 = conv_layers(128, max_pool_5)
conv_trans_1 = Conv3DTranspose(32, **conv_t_kwargs)(conv_6)
upconv_1 = conv_layers(64, concatenate([conv_5, conv_trans_1]))
conv_trans_2 = Conv3DTranspose(16, **conv_t_kwargs)(upconv_1)
upconv_2 = conv_layers(32, concatenate([conv_4, conv_trans_2]))
conv_trans_3 = Conv3DTranspose(8, **conv_t_kwargs)(upconv_2)
upconv_3 = conv_layers(16, concatenate([conv_3, conv_trans_3]))
conv_trans_4 = Conv3DTranspose(4, **conv_t_kwargs)(upconv_3)
upconv_4 = conv_layers(8, concatenate([conv_2, conv_trans_4]))
conv_trans_5 = Conv3DTranspose(2, **conv_t_kwargs)(upconv_4)
upconv_5 = conv_layers(4, concatenate([conv_1, conv_trans_5]))
output = Conv3D(1, kernel_size=1, activation='sigmoid')(upconv_5)
expected_model = Model(inputs=input_layer, outputs=output)
assert check_same_models(model, expected_model)
